AbstractMultilevel optimisation (MLO) refers to increasing the level of detail present in the optimisation problem through a decomposed set of sub-problems coordinated though some governing strategy. Within engineering MLO can be used to improve the preliminary structural design process by accounting for additional detail; increasing the level of design fidelity earlier allows for a greater range of nuanced solutions, improved robustness and greater accuracy in fundamental design decisions.
This research focused on practical considerations of MLO, applying algorithm from the major subsets of research identified within the literature to an industry guided design problem. This involved developing a generic modular framework to explore integrating detailed structural analyses for sub-components within a global wing sizing phase to minimise structural mass. Each algorithm was tested using varying assumptions whilst gradually increasing the difficulty of the design problem.
Monolithic methods were not robust enough for the most challenging problems but produced viable design solutions orders of magnitude more efficient than other methods tested. Dual distributed methods were the most reliable but overlapping algorithms, the breadth of techniques, tuning parameters and partitioning considerations means it can be challenging to select and tune the appropriate algorithm to produce useful results. Of the optimisers tested the most robust was a dual distributed method, augmented analytical target cascading with gradient-based partitioning.
Both monolithic and distributed MLO algorithm led to mass reductions of between 4-7% compared to traditional single-level design. Practical considerations for the application of MLO methods in industry include measuring the meaningfulness of decomposition to ensure sub-problems capture all the information of the original non-parsed problem. Additionally, pessimism can be used as a measure of competition or adversity between the system-subsystem problems, quantifiable through a range of factors which can significantly impact the applicability of any given method.
|Date of Award||28 Nov 2019|
|Supervisor||Alberto Pirrera (Supervisor) & Jonathan E Cooper (Supervisor)|